Medical documentation is often fragmented and error-prone, leading to compliance challenges and clinical risks. Guardoc Health aims to improve this by using Amazon Nova models to extract and classify information from complex documents. The solution helps skilled nursing facilities and assisted living centers process documents faster and more accurately than manual methods. This approach reduces the hundreds of millions of dollars spent annually on manual document processing in the U.S. According to Guardoc, the system achieves a 46% reduction in documentation errors, 70% fewer audit fines, and over $400K in annual return on investment for a single facility. Medical documents come in many formats, including multi-page PDFs with handwritten annotations, prior authorization forms, and medication lists embedded in tables or scanned images. Research shows that diagnostic errors affect about 12 million U.S. adults each year, with information-handling failures identified as a contributing factor. Guardoc Health was built to address these challenges by handling large volumes of documents with high precision. On peak days, they process over 1 million documents, and even a 1% error rate in condition detection generates thousands of incorrect records daily. This requires solving three major challenges: detecting special medical conditions with high recall, interpreting PDF checkboxes across form types, and extracting information from mixed-format documents. The solution involves using Amazon Nova models on Amazon Bedrock, combining several AWS services to handle each stage of document processing. Medical condition classification uses Retrieval Augmented Generation (RAG), which retrieves evidence from patient documentation and reasons across it to produce a final classification. The pipeline starts with Amazon Textract, which extracts text and structural metadata from each page. The extracted content is chunked into meaningful clinical units, such as medication lists or physician notes, to ensure accurate classification. Each chunk is embedded using Amazon Titan Text Embeddings V2 and stored in Amazon DynamoDB, partitioned by patient to avoid cross-patient retrieval. A custom pre-filter narrows the candidate set by evaluating document type, recency, and patient context, reducing the volume of content that needs to be searched. An in-memory k-nearest neighbor (k-NN) search retrieves the most relevant chunks for condition classification queries. Amazon Nova 2 Lite is used for lightweight, cost-efficient processing, while Amazon Nova Pro handles the final multimodal reasoning over raw PDF bytes. Every classification produced by the pipeline traces back to a specific source page in the original document. Guardoc Health treats this level of traceability as non-negotiable in clinical settings where decisions carry direct patient safety implications. The company operates multiple document processing pipelines, each tailored to address specific clinical challenges in extraction, classification, and search. The system follows a cost-tiering principle, using the least expensive component capable of the work at each stage. Amazon Titan Text Embeddings and Nova 2 Lite handle high-volume, lightweight tasks, while Nova Pro is reserved for the final stage requiring multimodal reasoning over raw PDF bytes. The pipeline ensures that every classification resolves back to a specific source page in the original document. Guardoc Health uses the Amazon Nova family of models and Amazon Bedrock to transform clinical documentation in long-term care. The solution helps healthcare organizations reduce the hundreds of millions of dollars spent annually on manual document processing across the United States. Guardoc reports a 46 percent reduction in documentation errors, 70 percent fewer audit fines, and over $400K in annual return on investment (ROI) for a single facility. Medical documents don’t follow rules, arriving in every format imaginable. A healthcare organization might process multi-page PDFs with handwritten annotations alongside printed text, prior authorization forms where checkbox states determine coverage decisions, medication lists embedded in tables, free text, or scanned images, and patient intake forms combining typed fields, handwriting, and rubber stamps. According to research published in BMJ Quality and Safety, diagnostic errors affect approximately 12 million U.S. adults each year in outpatient care, with information-handling failures identified as a contributing factor. Guardoc Health was built to address this variable. For Guardoc Health, the challenge is scale combined with precision. On peak days, they process more than 1 million documents. At that volume, even a 1 percent error rate in condition detection generates thousands of incorrect records daily, each one a potential patient safety risk or compliance liability. Achieving this required solving three of the hardest challenges in clinical document processing: detecting special medical conditions from patient records with high recall, reliably interpreting PDF checkboxes across dozens of form types, and accurately extracting information from documents that mix formats within a single page. The solution: Amazon Nova models on Amazon Bedrock. Guardoc Health built its pipeline on the Amazon Nova family of models and Amazon Bedrock, combining several AWS services to handle each stage of document processing. The following sections describe how each part works. Medical condition classification with RAG. Guardoc Health uses Retrieval Augmented Generation (RAG) to identify medical conditions from patient records. RAG is a technique in which a model first retrieves relevant information from an external source, then uses that context to generate a more accurate response. In this pipeline, Guardoc retrieves evidence directly from a patient’s own documentation and reasons across it to produce a final classification. Figure 1 — The Guardoc retrieval-augmented generation pipeline. As shown in the preceding diagram, the pipeline runs in the following order: Incoming patient documents pass through Amazon Textract, which extracts text and structural metadata from each page. This is the first stage of the pipeline for a comprehensive extraction at a low per-page cost. Amazon Textract is an AWS machine learning (ML) service that automatically extracts text, handwriting, forms, tables, and layout information from scanned documents and PDFs. It goes beyond basic optical character recognition (OCR) by identifying structure, so the output is usable in downstream apps and document workflows. The extracted content is chunked along boundaries appropriate to the document type, so chunks correspond to meaningful clinical units rather than arbitrary byte ranges. Chunks are smaller segments of the document aligned to meaningful clinical boundaries such as a medication list, a diagnosis section, or a physician note, rather than split arbitrarily by character count or byte size. This makes sure that when the pipeline searches for relevant information, it retrieves complete, coherent clinical units rather than fragments that lack the context needed for accurate classification. Each chunk is embedded using Amazon Titan Text Embeddings V2 and stored in Amazon DynamoDB, partitioned by patient so retrieval doesn’t cross patient boundaries. Before retrieval begins, a custom pre-filter narrows the candidate set by evaluating document type, recency, and patient-context signals. This step removes irrelevant documents early in the pipeline, reducing the volume of content that needs to be searched and lowering the overall cost of retrieval. Within the pre-filtered set, an in-memory k-nearest neighbor (k-NN) search retrieves the chunks most relevant to the condition classification query. k-NN is a technique that identifies the most similar items by measuring the distance between vector representations, returning the closest matches to the query. Because retrieval is scoped per patient, the system scales horizontally with patient volume without relying on a shared vector index. This stage returns page references only, keeping data transfer lightweight before the more intensive multimodal step that follows. The pages identified in the previous step pass through Amazon Nova 2 Lite, a lightweight, cost-efficient model that performs a text-based review to remove obvious non-matches before the pipeline advances to the more computationally intensive multimodal step. Amazon Nova 2 Lite is a low-cost, multimodal foundation model (FM) in the Amazon Nova family, designed for fast, affordable processing of text, images, and video inputs. The pages that survive all prior stages reach Amazon Nova Pro along with the raw PDF bytes. The Nova model reasons over layout, handwriting, signatures, stamps, and other visual context to produce the final classification, and every output resolves back to the specific source page it came from. Every classification produced by the pipeline traces back to a specific source page in the original document. In a clinical setting, where decisions carry direct patient safety implications, Guardoc Health treats this level of traceability as non-negotiable. Guardoc operates multiple document processing pipelines, each purpose-built to address a specific clinical challenge across extraction, classification, and search, as described in the following sections. PDF processing with Amazon Nova family of models. Most medical documents exist natively as PDFs, whether scanned, photographed, or born-digital, and processing them accurately requires a model that reads pages visually rather than as plain text. During evaluation, Amazon Nova Pro demonstrated a meaningful improvement in multimodal PDF handling, including layout interpretation, form structure, checkbox states, signatures, handwritten
Source: awsml
